replicating experiments

[dribnet]

Hi - I was hoping this repo would be suitable for replicating some of the paper's experiments, but it appears the code here is currently incomplete - the data loading portions including data.abstract.corpus are dependencies, but are not currently included. Is this correct, or is there something I'm missing?

[jacobandreas]

No, I just forgot to check in all the data and task loading code. I'm cleaning these up and writing instructions for running the experiments---should have this all ready soon.

[jacobandreas]

OK, should be good to go for colors and driving---if you need birds I can send you the file directly. I'm still trying to pin down a couple of nondeterminisims in the training code, so numbers may not be exactly the same. Qualitative behavior is stable.

[dribnet]

Thanks, confirmed. I'm now able to run the color_train config. I had to make a few changes to get it run in python3, which I've offered in #2.

Here's some notes on my process that could help others:

• installed pyyaml dependency
• commented out import cairo in color.py (only needed for visualize)
• follow README instructions for getting munroecorpus and also mkdir experiments
• ./run.sh configs/color_train.yaml (perhaps after tweaking run.sh)
• if there is an indent error in munroecorpus, just fix the one offending line - replacing spaces with a tab
• results/log in experiments/color_train. be sure to check run.log for errors